hako-mikan/sd-webui-supermerger

What it solves

SuperMerger is an extension for AUTOMATIC1111's Stable Diffusion WebUI that streamlines the process of merging AI image generation models. It eliminates the need to save and reload models to disk (HDD/SSD) during the merging process, allowing users to test different merge ratios and configurations in real-time without wasting storage space or time.

How it works

The tool allows users to merge multiple models (A, B, and C) using various calculation methods (calcmode) and merge modes (such as Weight Sum or Add Difference). A key feature is Merge Block Weight (MBW), which allows users to set different merge ratios for specific blocks of the model architecture (e.g., the U-Net's input and output blocks), enabling precise control over elements like art style, characters, and background.

It also includes an XYZ Plot for sequential generation, allowing users to visualize how different alphas, betas, or block weights affect the output across a grid of images. Additionally, it provides tools for adjusting detail, noise, and color tones directly within the model's U-Net without using LoRAs.

Who it’s for

Artists and model creators who use Stable Diffusion and want to experiment with model merging to create custom hybrid models with high precision and efficiency.

Highlights

• In-Memory Merging: Load merged models directly for generation without saving them to disk.
• Block-by-Block Control: Set specific merge ratios for each block of the model (SD 1.X, 2.X, XL, Flux, and Forge NEO DiT).
• Advanced Merge Modes: Supports Weight Sum, Add Difference, Triple Sum, and Sum Twice.
• Real-time Visualization: Use XYZ Plots to test multiple merge parameters and seeds simultaneously.
• Model Adjustment: Fine-tune detail, contrast, brightness, and color tones via the Adjust tab.
• LoRA Integration: Merge LoRAs, merge LoRAs into checkpoints, and extract LoRAs from checkpoints.